You will be able to point out machine learning in everyday apps and say what each one learned from.
Open the photo app on your phone and look at how it sorts your pictures. You probably never labelled any of them yourself, yet the app may already have picked out scenes, people and pets. Apple, for example, says its Photos app uses on-device machine learning for scene classification and for identifying people and pets. That sorting comes from machine learning.
Think through an ordinary morning. Your phone opens when you look at your face on the screen. On the MRT, the first video YouTube shows you is from a channel you watched once last week. At lunch you open your photos and see that the app has already picked out the people and pets in them, even though you never labelled any. Later your bank sends an SMS asking you to confirm a payment you did make, at a shop you rarely visit. None of these apps calls itself AI on screen, but all of them use machine learning. Once you can spot it in apps you already use, you can ask better questions about any product that claims to be AI-powered.
Netflix rows, the next video on YouTube and the suggested items under a product on Shopee all come from recommendation systems. These systems learn from behaviour. Every click is an example. So is every video you watch to the end, and every item you add to your cart and then abandon.
They also learn from other people. If many people who bought a rice cooker also bought a particular steamer tray, the system learns that link and shows the tray to the next person who buys a rice cooker. A feed can feel uncanny because it matches you with people whose habits look like yours.
You've probably seen how this goes wrong. Say you watch two cooking videos to help a friend, and your feed fills with recipes for a week. The system recorded the click but has no record of why you clicked. Recommenders learn from the past, so they are also slow to notice when your interests change.
These three features run on deep learning, the many-layered neural networks from Lesson 1.1. They handle the "fuzzy" kind of task from Lesson 1.2. Nobody can write rules for what makes your face your face in dim light with your glasses off, or for what separates a receipt from a menu in a photo.
So these models were trained on very large numbers of images and recordings. A photo search model learned from images paired with descriptions of what they contain. A speech model learned from audio paired with the words spoken. The app then uses the trained model that came out of that process. Where that model runs differs by app, and Apple says its Photos analysis of scenes, people and pets happens on the device itself.
When these models fail, the failures follow the training data. Voice typing does well on accents and words that were common in its training audio and worse on everything else. It is likely to guess wrong on Singlish, on the names of hawker centres, or on a sentence that switches between English and Mandarin. Face recognition can struggle in lighting it rarely saw during training. When one of these features slips, look first at what the model learned from.
A bank fraud alert and an email spam filter are both classifiers. A classifier is a model that sorts each new case into one of a few groups, such as fraud or not fraud, or spam or not spam. It is trained on past cases that people have already labelled. When a customer reports a transaction as fraud, or you mark an email as spam, that becomes a labelled example the system can learn from later.
Inside, a classifier usually produces a score instead of a plain yes or no. For a payment, the score is how likely it is to be fraud. The bank then chooses a cut-off, the threshold that turns the score into a decision such as sending an alert. A low cut-off gives more alerts, including false ones like the SMS about a payment you really made. A high cut-off bothers genuine customers less, but more fraud slips through. A person at the bank decided where the cut-off sits, and the model only supplies the score.
The label "AI-powered" appears on everything from aircons to CV screening software. On its own it tells you very little. It could mean a deep learning model, a simple classifier, or hand-written rules with new marketing. Ask two questions of any product that carries it.
First, ask what it learned from. A CV screening tool trained on a company's past hiring decisions copies that company's past preferences, both good and bad. A recommendation engine trained mostly on shoppers in another country may not suit people in Singapore. If the seller cannot tell you what the data was, treat that as your answer. Lesson 2.3, "A model can only be as good as its examples", comes next on this theme and covers this question in more depth.
Second, ask what it does when it is unsure. A well-built system admits doubt by passing the case to a person, asking you to confirm, or showing how confident it is. A weaker system guesses and gives you no sign that it guessed. When your bank asks you to confirm a payment, that is the well-built version, even if it is mildly annoying when you're standing at the checkout.
Your phone is the easiest place to practise, because it is full of features that quietly learned from somebody's data. Start noticing where these models are hiding in it. The activity asks you to find a few of them.
Open three apps on your phone. In each, write down one feature that probably uses machine learning and the data it probably learned from.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).